The agent mode isn't as user-friendly as Trae or Cursor

[5bug]

Validations

• [ ] I believe this is a way to improve. I'll try to join the Continue Discord for questions
• [ ] I'm not able to find an open issue that requests the same enhancement

Problem

The agent mode isn't as user-friendly as Trae or Cursor. Are there any plans to update and improve it?

Solution

No response

[PlebeiusGaragicus]

Came here for this.

I cannot get "agent" mode to edit any files.

I've been using a remote Ollama instance with Qwen 2.5 Coder (32b) and it seems to not work at all. I've also tried QwQ and now trying to get DeepSeek Coder V2.

I can't figure out if I'm confused on how you're supposed to use it... but this feature doesn't even seem ready to ship. I have NOT tried with with Claude, or Gemini, though. So perhaps my ollama models just aren't good enough(?)

I use Windsurf and it's AMAZING how great their agent mode is - I am still hoping to switch to Continue... but I literally can't get any LLM agent to use a tool properly.

[elijahr]

I am having a similar struggle. I have tried a variety of models on a remote Ollama, (qwen2.5-coder:latest, codestral, deepseek-coder, incept5/llama3.1-claude). For those that I am able to select "Agent" mode, it behaves very differently than I am expecting.

Some of the differences I chalk up to the fact that the open models I have tried are not as good as the proprietary models, such as Gemini. But the big issue I have, regardless of the model, and this seems like a bug, is that instead of actually making any changes to files, Continue just prints some code to the chat pane for me to save manually.

When I ask "Why are you not able to edit for me?", I get messages along the lines of: "I'm sorry for any confusion, but as an AI model, I don't have direct access to file systems or the ability to edit files on your local machine or in a repository. My primary function is to provide information and assistance through text-based interactions."

Am I doing something wrong? Is my expectation that Continue is supposed to be able to make edits incorrect?

[5bug]

Looking forward to continue achieving an interactive experience similar to cursor/trae.

[chezsmithy]

https://github.com/continuedev/continue/blob/main/core/llm/toolSupport.ts#L83

Please validate your model is in the list of tool supported models. It looks possible it might be, but please validate.

If not, please choose a supported model. It may be possible to enable support for specific models with minor changes.

@RomneyDa is it possible agent mode should be disabled if tools support is not available from selected models?

[son1ss]

I think it may be related to that <important-rules> section

[github-actions[bot]]

This issue hasn't been updated in 90 days and will be closed after an additional 10 days without activity. If it's still important, please leave a comment and share any new information that would help us address the issue.

[mxmlnkn]

Please don't auto-close this issue. It is unspecific, yes, but I have had the same user experience. I want to switch away from Cursors, which worked amazingly well.

• First, I had this problem: https://github.com/continuedev/continue/issues/5182. Easily worked around with: https://github.com/continuedev/continue/issues/5182#issuecomment-3072205195
• The next problem I encountered was that changing a model in the request may not change the model for other actions, such as edit / apply / agent, ... This resulted in the chat working but in timeouts when trying to apply a change because it tried to request a model that was not loaded. I know, "user error", but maybe show a faster warning. It takes quite long to time out. Maybe it could communicate with LM Studio and let it tell the actually loaded models instead of all available ones. And an option to switch all action models at once would also be helpful.
• On startup, the "Indexing" tab always shows "Cannot read properties of undefined (reading 'signal')", but when clicking "retry", it always works, so it makes no sense.
• The default indexing takes so long that it is unusable because it tries to index committed .dat and .bz2 files. I feel like there should be some better defaults, e.g., a default cursorignore and maybe something better than file extension matching, e.g., magic bytes matching. How is binary data, such as bzip2 even tokenized?
• Then, I ran into context length too small errors. I am not sure why it would error out on this instead of simply letting it truncate, which seems more user-friendly, even if it would lead to less good results. It might have been another problem with the AUTODETECT LM Studio setup. Maybe it does not communicate the context length correctly. After increasing it to 24K+ it began to work somewhat. 16K also resulted in errors because it was too short, depending on the query.
• It feels like the indexing has limited use. Once, I tried to let it refactor my code base and it began by websearching for the codebase's internal functions like a nooby developer, which of course will not yield useful results. This was even though I have added @Codebase, and the indexing was successful. I have since enabled the experimental options to always add the codebase and current file, which seems like a good default if the model context length allows it.
• With Qwen QWQ 32B, which runs on my CPU, I got the first reasonable results comparable to my Cursor experience. It is too slow for actual usage, but ok for simply trying out what is possible. However, even there, I get a "Search truncation warning". I think I got something similar in Cursor, though. I would hope that the next step would have been to continue the search and then edit those files, and repeat until all search results have been processed. Unfortunately, I did not try out the next step because editing the files seems to have failed for some unknown reason: That last message is weirdly what led me to this issue via the search, even though it is not mentioned here. If it tries to tell me that it can only edit files that are opened in the editor, then why not simply open them automatically, like Cursor did?

[ShaunaGordon]

I am having a similar struggle. I have tried a variety of models on a remote Ollama, (qwen2.5-coder:latest, codestral, deepseek-coder, incept5/llama3.1-claude). For those that I am able to select "Agent" mode, it behaves very differently than I am expecting.

Some of the differences I chalk up to the fact that the open models I have tried are not as good as the proprietary models, such as Gemini. But the big issue I have, regardless of the model, and this seems like a bug, is that instead of actually making any changes to files, Continue just prints some code to the chat pane for me to save manually.

When I ask "Why are you not able to edit for me?", I get messages along the lines of: "I'm sorry for any confusion, but as an AI model, I don't have direct access to file systems or the ability to edit files on your local machine or in a repository. My primary function is to provide information and assistance through text-based interactions."

Am I doing something wrong? Is my expectation that Continue is supposed to be able to make edits incorrect?

Continue, by default, asks for confirmation on everything. This is an intentional decision, as I understand it, favoring safety. You can click on the gear icon on the Continue panel and under Experimental Settings, enable "Auto-accept Agent Edits".

[son1ss]

Continue, by default, asks for confirmation on everything. This is an intentional decision, as I understand it, favoring safety. You can click on the gear icon on the Continue panel and under Experimental Settings, enable "Auto-accept Agent Edits".

I think the problem described is not the confirmation dialog, but rather that the model behaves like in regular chat mode instead of being an agent and asking and actively doing stuff

I've had encountered this myself with qwen2.5-coder:14b

[ShaunaGordon]

Continue, by default, asks for confirmation on everything. This is an intentional decision, as I understand it, favoring safety. You can click on the gear icon on the Continue panel and under Experimental Settings, enable "Auto-accept Agent Edits".

I think the problem described is not the confirmation dialog, but rather that the model behaves like in regular chat mode instead of being an agent and asking and actively doing stuff

I've had encountered this myself with qwen2.5-coder:14b

I've seen it before, too, and there's usually been a confirmation button at the top of each block that will do the action. The times where I've seen it go full "chat" mode is when I've tried using models with questionable Tool capability.

[github-actions[bot]]

This issue hasn't been updated in 90 days and will be closed after an additional 10 days without activity. If it's still important, please leave a comment and share any new information that would help us address the issue.

[github-actions[bot]]

This issue was closed because it wasn't updated for 10 days after being marked stale. If it's still important, please reopen + comment and we'll gladly take another look!